Stop flattening EXIF IFD into getexif() #4947

radarhere · 2020-10-05T09:36:30Z

At the moment, the Exif class flattens the EXIF IFD into the rest of the Exif data. Losing information in this way is not ideal.

Here is the new test added to demonstrate the difference.

def test_exif_ifd(self):
    im = Image.open("Tests/images/flower.jpg")
    exif = im.getexif()

    reloaded_exif = Image.Exif()
    reloaded_exif.load(exif.tobytes())
    assert reloaded_exif.get_ifd(0x8769) == exif.get_ifd(0x8769)

To try and prevent this, while maintaining as much backwards compatibility as possible, this PR stops this flattening only for getexif(), while keeping it in place for the older _getexif(). So this is backwards incompatible, but only for code written after the release of 6.0.0.

Only one existing test had to be changed to match this breaking change.

kkopachev · 2020-10-06T19:37:45Z

src/PIL/Image.py

@@ -3346,91 +3351,100 @@ def tobytes(self, offset=8):
            head = b"MM\x00\x2A\x00\x00\x00\x08"
        ifd = TiffImagePlugin.ImageFileDirectory_v2(ifh=head)
        for tag, value in self.items():
+            if tag in [0x8769, 0x8225] and not isinstance(value, dict):
+                value = self.get_ifd(tag)


This does not work for interop since it's inside of Exif IFD.

im = Image.open('10_years_of_Wikipedia_by_Guillaume_Paumier.jpg') good_exif = im.getexif() bad_exif = Exif() bad_exif.load(good_exif.tobytes()) print("Good interop dictionary size: %d" % len(good_exif.get_ifd(0xA005))) print(" Bad interop dictionary size: %d" % len(bad_exif.get_ifd(0xA005)))

I suspect same for makernote too, but since makernote is vendor specific, I think it would be safer to make it read only property.

Ok, I've added another commit to cover Interop.

I've being thinking through this on and off for some time.
Thinking of consumer of this class, it would make sense when reading Exif IFD to get Interop value decoded into a dict at that time. Or shimmed with ImageFileDirectory_v2 instance. Otherwise caller must know Interop tag ID, must know to call get_ifd, etc. So that's a bit inconvenient and adds extra knowledge to consumer.
Same for Makernote, however saving it back we don't want to encode it back to vendor-specific format and probably worth just copying whatever bytes were there originally.
On the other hand Exif is quite complicated and some knowledge from consumer is expected anyway.
🤷
What do you think of this?

Another benefit of using ImageFileDirectory_v2 as a shim for sub-IFDs is that we get lazy loading.

kkopachev · 2020-10-06T20:05:18Z

Also consider that Exif might contain multiple IFDs one after another. Currently only first IFD is preserved on save.

src/PIL/Image.py

radarhere · 2020-10-07T12:36:58Z

Also consider that Exif might contain multiple IFDs one after another. Currently only first IFD is preserved on save.

I'm not following. What do you mean, 'only first IFD'?

kkopachev · 2020-10-07T17:19:39Z

I was referring to IFD1 - thumbnail image info which might be present in the file. Accessible through self._info.next after reading IFD0.
Not sure how important it is to preserve it, thought we're here already and it might affect code design anyway.

src/PIL/JpegImagePlugin.py

kkopachev · 2020-10-07T17:50:02Z

src/PIL/Image.py

-                                f"Expecting to read {size} bytes but only got "
-                                f"{len(data)}. Skipping tag {ifd_tag}"
+        if tag not in self._ifds:
+            if tag in [0x8769, 0x8825]:


I think __getitem__ method should be updated as well to call get_ifd for Exif IFD. I can't comment on non-changed code, but currently it's line 3474 should be if tag in [0x8769, 0x8825]:

While I can see that this would be more useful for most cases, if you wanted to actually read the IFD offset, you wouldn't be able to anymore (apart from looking directly into _data or _info)

I see your point.
I think getting offset of Exif IFD is useless by itself, unless you want manually parse IFD structure yourself, in which case you could do so with raw exif bytes.
Additionally, I think code should be consistent: either decode both GPS and Exif IFDs into dictionaries on read in __getitem__, or return offset in both cases.

radarhere · 2020-10-12T08:19:45Z

I was referring to IFD1 - thumbnail image info which might be present in the file. Accessible through self._info.next after reading IFD0.
Not sure how important it is to preserve it, thought we're here already and it might affect code design anyway.

Couldn't that easily lead to a situation where the thumbnail image no longer matches the main image, if the main image is modified by Pillow, but the thumbnail image isn't?

kkopachev · 2020-10-12T21:51:51Z

Couldn't that easily lead to a situation where the thumbnail image no longer matches the main image, if the main image is modified by Pillow, but the thumbnail image isn't?

That's a great point. In that case it's better to ignore IFD1 for writing.

Co-authored-by: Konstantin Kopachev <kkopachev@popsugar.com>

hugovk · 2021-03-07T10:29:29Z

@kkopachev How does this PR look now?

kkopachev · 2021-03-09T01:27:45Z

@hugovk Apart from inconsistent treatment of GPS IFD in __getitem__, it looks good

radarhere · 2021-03-15T02:41:16Z

Ok, I've added a commit to treat it consistently.

hugovk · 2021-03-19T19:26:52Z

Thanks both!

Added release notes for #4947

radarhere added the Exif label Oct 5, 2020

radarhere mentioned this pull request Oct 5, 2020

exif_transpose errors out on some images #3973

Closed

kkopachev reviewed Oct 6, 2020

View reviewed changes

kkopachev reviewed Oct 7, 2020

View reviewed changes

src/PIL/Image.py Outdated Show resolved Hide resolved

radarhere force-pushed the exif branch from d2c0a5a to b415b04 Compare October 7, 2020 08:04

kkopachev reviewed Oct 7, 2020

View reviewed changes

src/PIL/JpegImagePlugin.py Show resolved Hide resolved

kkopachev reviewed Oct 7, 2020

View reviewed changes

hugovk force-pushed the master branch from 747e44b to 9e3ad5e Compare December 18, 2020 20:12

radarhere force-pushed the exif branch from 3b571f4 to 8723cb3 Compare January 4, 2021 12:40

radarhere mentioned this pull request Feb 18, 2021

Pillow corrupts EXIF after changing it #5273

Closed

radarhere and others added 4 commits February 22, 2021 07:47

Stopped flattening EXIF IFD into getexif()

faf8fad

Save base IFDs when converting Exif to bytes

4b14f01

Simplified code

b25bc40

Co-authored-by: Konstantin Kopachev <kkopachev@popsugar.com>

Save interop IFD when converting Exif to bytes

e763f8f

radarhere force-pushed the exif branch from 8723cb3 to e763f8f Compare February 21, 2021 20:51

radarhere and others added 3 commits March 15, 2021 12:33

Removed automatic retrieval of GPS IFD

c52b45d

Merge branch 'master' into exif

68719fe

Updated comments

36a4b05

kkopachev approved these changes Mar 15, 2021

View reviewed changes

hugovk merged commit 5a20908 into python-pillow:master Mar 19, 2021

radarhere deleted the exif branch March 19, 2021 21:19

radarhere added a commit to radarhere/Pillow that referenced this pull request Mar 19, 2021

Document python-pillow#4947 [ci skip]

da9b104

hugovk added a commit that referenced this pull request Mar 19, 2021

Merge pull request #5344 from radarhere/releasenotes

e9c4834

Added release notes for #4947

radarhere mentioned this pull request Mar 29, 2021

EXIF inconsistencies for "UserComment" tag #5254

Closed

radarhere mentioned this pull request Jun 14, 2021

Return incorrect DPI #2817

Closed

radarhere mentioned this pull request Dec 2, 2021

How to get GPSInfo and DateTimeOriginal from .getexif()? #5863

Closed

radarhere mentioned this pull request Dec 17, 2021

getexif() returns incomplete exif data since 8.2.0 #5894

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Stop flattening EXIF IFD into getexif() #4947

Stop flattening EXIF IFD into getexif() #4947

radarhere commented Oct 5, 2020 •

edited

Loading

kkopachev Oct 6, 2020 •

edited by radarhere

Loading

radarhere Oct 7, 2020

kkopachev Oct 7, 2020 •

edited by radarhere

Loading

kkopachev Oct 7, 2020

kkopachev commented Oct 6, 2020

radarhere commented Oct 7, 2020

kkopachev commented Oct 7, 2020

kkopachev Oct 7, 2020

radarhere Oct 11, 2020

kkopachev Oct 12, 2020

radarhere commented Oct 12, 2020

kkopachev commented Oct 12, 2020

hugovk commented Mar 7, 2021

kkopachev commented Mar 9, 2021

radarhere commented Mar 15, 2021

hugovk commented Mar 19, 2021

Stop flattening EXIF IFD into getexif() #4947

Stop flattening EXIF IFD into getexif() #4947

Conversation

radarhere commented Oct 5, 2020 • edited Loading

kkopachev Oct 6, 2020 • edited by radarhere Loading

Choose a reason for hiding this comment

radarhere Oct 7, 2020

Choose a reason for hiding this comment

kkopachev Oct 7, 2020 • edited by radarhere Loading

Choose a reason for hiding this comment

kkopachev Oct 7, 2020

Choose a reason for hiding this comment

kkopachev commented Oct 6, 2020

radarhere commented Oct 7, 2020

kkopachev commented Oct 7, 2020

kkopachev Oct 7, 2020

Choose a reason for hiding this comment

radarhere Oct 11, 2020

Choose a reason for hiding this comment

kkopachev Oct 12, 2020

Choose a reason for hiding this comment

radarhere commented Oct 12, 2020

kkopachev commented Oct 12, 2020

hugovk commented Mar 7, 2021

kkopachev commented Mar 9, 2021

radarhere commented Mar 15, 2021

hugovk commented Mar 19, 2021

radarhere commented Oct 5, 2020 •

edited

Loading

kkopachev Oct 6, 2020 •

edited by radarhere

Loading

kkopachev Oct 7, 2020 •

edited by radarhere

Loading